How to train additional parameters

autograd
1

Hi, I am trying to implement a trainable parameter in an existing codebase, which is overwhelmingly large for me at the moment.

The original codebase can be found at https://github.com/TRI-ML/packnet-sfm

My main contribution should be trainable intrinsic parameters (which is basically a 3x3 matrix).

My best attempt is the following: a new Parameter is added to the SelfSupModel class which is set to zero in __init__ and the initial value is set at the first call of forward():

class SelfSupModel(SfmModel):
    def __init__(self, **kwargs):
        ...
        self.intrinsic = torch.nn.Parameter(torch.zeros((3, 3)))
        self.intrinsic.requires_grad = True
    ...
    def forward(self, batch, return_logs = False, progress = 0.0):
        ...
        if self.intrinsic.sum() == 0:
            self.intrinsic = torch.nn.Parameter(batch['intrinsics'][0])

        stored_intrinsic = torch.stack(batch['intrinsics'].shape[0] * [self.intrinsic], 0)
        self_sup_output = self.self_supervised_loss(
            batch['rgb_original'], batch['rgb_context_original'],
            output['inv_depths'], output['poses'], stored_intrinsic,
            return_logs = return_logs, progress = progress)

I set requires_grad = True and the intrinsic can be found in SelfSupModel.named_parameters() with requires_grad = True.

Upon printing out this Parameter, it is not updated after each epoch.
What can be the reason for not being updated?

I assume that this is not the best possible way to initialize this Parameter and this might not be optimized for GPU.

Thanks, Fabian

2

Sample-specific values are not considered model parameters. Though you can train such values, you won’t be able to make out-of-sample predictions. Perhaps you need group-specific values. In both cases, you should probably use nn.Embedding index lookups (do embedding.weight.data.zero_() as it is initialized with gaussian noise by default)

3

Thank you Alex,
out-of-sample prediction might be useful later on, so I think I should add prediction of intrinsic parameters to the PoseNet.

Fabian